Build an AI Knowledge Base in Obsidian
Build an Obsidian vault that gives AI traceable sources, current decisions, and a clear way to handle unknowns.
An Obsidian vault is ready for AI-assisted work when someone can trace a conclusion back to the source and tell whether it still applies. Keep evidence, decisions, and approved publishing material in separate folders. Ask the model to cite its evidence before you use its answer.
What folders are enough to start?
| Folder | Contents | Edit rule |
|---|---|---|
| sources | Dated notes from authorized reports, observations, and documents | Preserve the original and add new records |
| decisions | Confirmed conclusions, scope, and review dates | A person approves the decision |
| publish | Material reviewed for rights, privacy, and public use | Review before publishing |
Start with one operating question, such as why a particular product has more returns this week. The evidence needs its date range and scope. A decision needs a review condition. A public draft needs only material you can use outside the team.
What properties should you keep on each note?
Track a source, observation date, scope, status, and next review date. Use statuses such as unverified, current, and superseded. When a decision changes, keep the old note and link to its replacement. Obsidian properties can store dates and links. Internal links connect a decision to the observations that support it. Obsidian Properties · Obsidian internal links
Copy this example into the top of a decision note and replace the source, dates, and scope with your own.
---
source: "[[sources/First week of November returns]]"
observed: 2026-11-07
scope: "Storage basket A in the US shop"
status: unverified
review: 2026-11-14
---
How can you test whether an AI answer is grounded?
Give the model the notes relevant to the question. Request an answer, source links, and any unknowns. Suppose an old return report mentions sizing and a newer one mentions color. Ask the model to check each product and period, then explain where the figures are comparable. It should leave unsupported numbers in the unknowns section. Keep source notes available to the reviewer and limit the input to the data needed for the task.
When is it ready for another operator?
Try three questions before expanding the library. Find a current rule and its evidence. Locate a rule that was superseded and show its replacement. Ask about a fact the vault does not contain. Then add a catalog and check links as the library grows. SCTCamp's open Vault Keeper repository illustrates this approach with vault rules, a generated catalog, and checks. Vault Keeper on GitHub